Kent Academic Repository

Full text document (pdf)

Citation for published version
Lin, Junyi and Spiegler, V.L.M. and Naim, M.M. (2018) Dynamic analysis and design of a semiconductor
supply chain: a control engineering approach. International Journal of Production Research,
56 (13). pp. 4585-4611. ISSN 0020-7543.

DOI
https://doi.org/10.1080/00207543.2017.1396507

Link to record in KAR

https://kar.kent.ac.uk/64241/

Document Version
Author's Accepted Manuscript

Copyright & reuse

Content in the Kent Academic Repository is made available for research purposes. Unless otherwise stated all
content is protected by copyright and in the absence of an open licence (eg Creative Commons), permissions
for further reuse of content should be sought from the publisher, author or other copyright holder.

Versions of research
The version in the Kent Academic Repository may differ from the final published version.
Users are advised to check http://kar.kent.ac.uk for the status of the paper. Users should always cite the
published version of record.

Enquiries
For any further enquiries regarding the licence status of this document, please contact:
researchsupport@kent.ac.uk
If you believe this document infringes copyright then please contact the KAR admin team with the take-down
information provided at http://kar.kent.ac.uk/contact.html
Dynamic analysis and design of a semiconductor supply chain: a control engineering
approach

Junyi Lin,
Logistics Systems Dynamics Group, Cardiff Business School, Cardiff University, Cardiff, UK.
linj17@cardiff.ac.uk

Virginia L.M. Spiegler,

Management Science Research Group, Kent Business School, University of Kent, Canterbury, UK.
v.l.spiegler@kent.ac.uk

M.M. Naim
Logistics Systems Dynamics Group, Cardiff Business School, Cardiff University, Cardiff, UK
naimmm@cardiff.ac.uk
Dynamic analysis and design of a semiconductor supply chain: a control
engineering approach
The combined make-to-stock and make-to-order (MTS-MTO) supply chain is well-recognized in the
semiconductor industry in order to find a competitive balance between agility, including customer
responsiveness, and minimum reasonable inventory, to achieve cost efficiency while maintaining customer
service levels. Such a hybrid MTS-MTO supply chain may suffer from the bullwhip effect, but few researchers
have attempted to understand the dynamic properties of such a hybrid system. We utilize a model of the Intel
supply chain to analytically explore the underlying mechanisms of bullwhip generation and compare its dynamic
performance to the well-known Inventory and Order Based Production Control System (IOBPCS) archetype.
Adopting a control engineering approach, we find that the feedforward forecasting compensation in the MTO
element plays a major role in the degree of bullwhip and the Customer Order Decoupling Point (CODP)
profoundly impacts both the bullwhip effect and the inventory variance in the MTS part. Thus, managers should
carefully tune the CODP inventory correction and balance the benefit between CODP inventory and bullwhip
costs in hybrid MTS-MTO supply chains.

Keywords: Make-to-Stock, Make-to-Order, hybrid MTS-MTO, semiconductor industry, IOBPCS family

1. Introduction
Facing a highly volatile and turbulent business environment, caused by reduced product life cycles
and the unpredictable and customized demand, agile supply chains have become a key capability for
current businesses to survive and thrive (Braunscheidel and Suresh 2009). On the other hand, the
pressure of leaner supply chains forces practitioners to focus on minimum reasonable inventory (MRI)
(Grünwald and Fortuin 1992) and on corresponding issues, such as forecasting accuracy and safety
stock (Dudek and Stadtler 2005; Gunasekaran, Patel, and McGaughey 2004).
This is the case in the capital-intensive semiconductor industry characterized by a short product
life cycle, wide product variety due to overlapping product life cycles for different customers, long
fabrication lead times and complex production processes (Geng and Jiang 2009). To find a competitive
balance between agility (customer responsiveness) and MRI (cost efficiency), the combination of
make-to-stock and make-to-order (MTS-MTO), also called the hybrid push-pull strategy, has been
considered in the semiconductor industry ( Lee et al. 2006; Hur, Bard and Chacon 2017). A hybrid
MTS-MTO approach refers to the supply chain strategy of postponing the customization of products
at different levels until the actual customer orders are received (Kim et al. 2012). The boundary
between MTO and MTS, called the ‘Customer Order Decoupling Point’ (CODP) (Naylor, Naim, and
Berry 1999; Harrison, Lee, and Neale 2005) or ‘order penetration point’ (Olhager 2003; Christopher
2016), is the point that separates the forecasting-based MTS element that pushes semi-finished
products, and the order-based MTO part that pulls products through for the customization process
(Chen and Dong 2017).
Although such a hybrid MTS-MTO strategy has been well-recognized in the semiconductor
industry, few studies have focused on understanding the hybrid MTS-MTO environment from a supply
chain dynamics perspective. Most literature on the modelling and analysis of the hybrid MTS-MTO
strategy in the semiconductor industry has explored scheduling and control (Chang et al. 2003; Wang,
Rivera, and Kempf 2007), the theoretical and empirical development of a hybrid MTS-MTO model
(Brown, Lee, and Petrakian 2000; Lee et al. 2006) and postponement analysis (Kim and Kim 2012).
Semiconductor firms have suffered severely from demand amplification, or the bullwhip effect
(Hofmann 2017; Li and Disney 2017; Vicente, Relvas and Barbosa-Póvoa 2017), due to their relatively
extreme distance in the entire supply chain from end customers (Brown, Lee, and Petrakian 2000), as
well as the characteristics of high levels of stochasticity and nonlinearity (Wang and Rivera 2008). For
this reason, we aim to explore analytically the underlying mechanisms of supply chain dynamics within
the context of the semiconductor industry. We use the supply chain model of Intel, the leader in
microprocessor manufacturing (Sampath et al. 2015), as reported empirically by Gonçalves, Hines,
and Sterman (2005), as a base framework to extract and analyse the MTS-MTO supply chain.
Moreover, we benchmark the MTS-MTO model’s dynamic behaviour with a well-established supply
chain family of model archetype, the inventory and order-based production control (IOBPCS) (Towill
1982; John, Naim, and Towill 1994; Sarimveis et al. 2008; Lin et al. 2016).
In doing so, we address the following research questions:

1. How can we gain insight into the dynamic properties of a semiconductor hybrid MTS-MTO supply
chain as personified by the Intel model?
2. What are the underlying mechanisms of the dynamic behaviour in a semiconductor hybrid MTS-
MTO supply chain and how can these dynamics be mitigated?

Figure 1 illustrates our research approach. We first model the Intel supply chain in a block diagram
form based on the model descriptions given by Gonçalves, Hines and Sterman (2005). Although their
model provided insights into lean inventory and responsive utilization policies, their simulation
approach is not able to reveal the explicit relationship between the system’s outputs and the
endogenous demand, therefore overlooking the real effects of some control parameters. After
simplifying the block diagram and extracting the MTS-MTO scenario (push-pull), we analyse the
dynamic behaviour of the system by finding the system’s transfer functions. The simplified model
enables us to draw an analogy with known archetypes of the IOBPCS family and to propose policies
to overcome trade-offs in the system output responses. Although we investigate the dynamic behaviour
of the hybrid MTS-MTO model within the Intel supply chain, the insights gained from the analytical
results can be generalized to other high-volume and low-variety semiconductor supply chains.

Based on Gonçalves,
Modelling Intel supply chains in a
Hines, and Sterman
block diagram form
(2005)

1. Eliminate nonlinearities
Simpify the Intel supply chains to 2. Eliminate redundancies
extract the MTS-MTO scenario 3. Collecting terms
4. Draw the analogy to the IOBPCS family

1. Laplace Transform
2. Characteristics equations analysis
Cross-check by Analyze the dynamic properties of the
3. Dynamic analysis in responding unit step
simulation MTS-MTO scenario
input
4. Compare with the IOBPCS family

Conclusion

Figure 1. The research approach.

2. Literature review
2.1 The hybrid MTS-MTO strategy in semiconductor production strategy
Although the complex material and equipment acquisition processes vary between different
companies, a typical semiconductor manufacturing process consists of three main stages: wafer
fabrication (‘front-end’ manufacturing), assembly and test, and product distribution (‘back-end’
operations), whose associated activities are usually involved in a globally-complex network to save
labour costs and benefit from tax breaks (Rastogi et al. 2011). Depending on the time horizon,
production planning and control in the semiconductor industry can be divided from yearly-based
planning, monthly/weekly-based order release to daily scheduling and hourly dispatching (Mönch,
Fowler and Mason 2013). As this study focuses on mid-/long-term production planning, we review the
main operational strategies adopted for such planning periods and order release procedures.
At a detailed machine level, extensive literature has studied the push, pull or hybrid order release
approaches for semiconductor wafer fabrication, in which the push strategy refers to the production
approach based on desired quantity of goods, while pull-based methods focus on what can be produced
based on real-time demand and resource constraints. The authors refer to; Chandra and Gupta (1997);
Bahaji and Kuhl (2008);Qi, Sivakumar, and Gershwin (2008); Lin and Lee (2011) and Zhang Bard and
Chacon (2017) for details.From a broader and more strategic supply chain operational perspective,
there are three strategies commonly considered by the semiconductor industry: MTS, MTO or ATO
(assembly-to-order, that is, a kind of hybrid MTS-MTO that places the decoupling point at the
assembly echelon) (Forstner and Mönch 2013; Sun et al. 2010). Utilizing a case study, Brown et al.
(2000) highlighted the benefit of implementing the hybrid MTS-MTO for better financial performance
via holding less finished inventory. Lee (2001) discussed the consequences of implementing a hybrid
production strategy with die bank as the decoupling point in the semiconductor context and
recommended a postponement strategy for operating such complex supply networks. Sun et al. (2010)
proposed a hierarchical decision support framework to guide the selection of MTS, MTO or ATO
strategies for designing semiconductor supply chains. The results indicated that the decision is driven
by two customer-oriented factors, i.e. required lead time and the importance of on-time delivery
performance, although customer demand patterns and process variability may play an important role
under certain circumstances. Forstner and Mönch (2013) developed a genetic algorithm (GA) to
support the selection of different production strategies (i.e. MTS, MTO, ATO) by using discrete event
simulation. A profit-based objective function is developed under the consideration of stochastic
behaviour of the semiconductor supply chain, although GA is criticized as a time-consuming procedure
to assess the fitness of the right production strategy.
While previous research has explored the possible benefits and selection criteria for different
supply chain operational strategies in the semiconductor industry, very few studies focus on the
dynamic behaviour of the hybrid MTS-MTO strategy, except for Gonçalves, Hines, and Sterman
(2005); Orcun, Uzsoy, and Kempf (2006) and Orcun and Uzsoy (2011). Specifically, Gonçalves, Hines,
and Sterman (2005) developed a system dynamics simulation model to explore how market sales and
production decisions interact to create unwanted production and inventory variances in the Intel hybrid
MTS-MTO supply chain. Using a system dynamics approach, Orcun, Uzsoy and Kempf (2006)
developed a capacitated semiconductor production model with load-dependent lead time, which
overcomes the limitation of treating lead times as exogenous parameters independent of the decision
variables that most linear dynamic models assume. The analysis suggested that the nonlinear change
at high capacity utilization is consistent with insights from queuing models and industrial practices.
Furthermore, Orçun and Uzsoy (2011) studied the dynamic behaviour of a simplified semiconductor
supply chain system with two capacitated manufacturing echelons and one inventory echelon. They
indicated that the dynamic properties of a supply chain system under optimization-based planning
models are qualitatively different from those operating under simple feedback policies system
dynamics models.
Although these system dynamics simulations contribute to the representation of a real system by
incorporating nonlinear components and complex structures, it is a trial-and-error approach that may
hinder the system improvement process (Towill 1982; Sarimveis et al. 2008). Despite the fact that
semiconductor supply chains have suffered severely from the bullwhip effect (Chien, Chen, and Peng
2010; Terwiesch et al. 2005), limited effort has been made to explore the underlying system structures
that cause the phenomenon. As a result, there is a need to consider analytical methods to understand
the underlying mechanisms of bullwhip generation and propose corresponding mitigation approaches
that are relevant for the semiconductor hybrid MTS-MTO supply chain.

2.2 Classic control theory and the IOBPCS family in studying supply chain dynamics
 Classic control theory techniques, with feedback thinking and sufficient analytical tools, are
advantageous for analysing supply chain dynamics (Sarimveis et al. 2008). The application of classic
control theory in a production system can be traced back to Simon (1952). Table 1 gives a brief
introduction of control engineering relevant tools/methods utilized in this study.

Tools/Methods Description and advantages References (e.g.)

Block diagram Block diagrams are used to outline a system in which the principal parts Schwarzenbach and Gill
or functions are represented by blocks connected by lines that show the (1992)
relationships of the blocks.
The Block diagrams are useful to describe the overall concept of a Disney and Towill
complex system without concerning the details of implementation, (2002); Dejonckheere et
which allow for both a visual and an analytical representation within a al. (2004); Spiegler et al.
single entity. The adoption of block diagrams in studying supply chain (2016)
dynamics has been well-recognized in production planning and control
literature.
Laplace The Laplace transform is an integral transform in which convert a Schwarzenbach and Gill
transformation function of a real variable t (time domain) to a function of a complex (1992)
variable s (frequency domain), shown as follow:

The Laplace transform technique has great advantages of simplifying Disney and Towill
the algebraic manipulations required, analysing large systems, handling (2002); Disney, Towill,
arbitrary inputs and benchmarking good practice in studying supply and Warburton (2006)
chain dynamics.
Transfer function The transfer function of a system is a mathematical representation Nise (2007)
describing the dynamic behaviour in a linear, time-invariant (LTI)
system algebraically. It can be defined as the ratio of s/z transform of
the output variables to the s/z-transform of the input variables,
depending on the consideration of variables change with time
continuously or discretely.

1. The transfer function approach can be used to model Dejonckheere et al.

production/supply chain systems, since they can be seen as systems (2003); Spiegler, Naim,
with complex interactions between different parts of the chain. and Wikner (2012)
2. Transfer functions completely represent the dynamic behaviour of
production/supply chain systems under a particular replenishment
rule, i.e. the input to the system represents a specific demand pattern
and the output refers to the corresponding replenishment or production
orders.
Table 1. A brief review of control engineering tools/methods utilized in this study.

By adopting classic control theory, Towill (1982) translated Coyle (1977) system dynamics
work to represent the IOBPCS in a block diagram form. John, Naim, and Towill (1994) then extended
the original model to the automatic pipeline, inventory and order-based production control system
(APIOBPCS) by incorporating an automatic work-in-progress feedback loop. These two original
models and their variants, i.e. the IOBPCS family, have been recognized as a base framework for
production planning and control systems, as they consist of general laws that represent many supply
chain contexts, such as the famous beer game decision-making heuristic (Sterman 1989), the order-up-
to (OUT) policy represented by APVIOBPCS (Automatic pipeline and various inventory and order-
based production control system) (Chen and Disney 2007; Zhou, Disney, and Towill 2010),
remanufacturing system (Zhou et al. 2006; Zhou, Naim, and Disney 2017) and various other industrial
applications (e.g. Coyle 1977). The authors refer to Lin et al. (2016) for a full review of the IOBPCS
family in studying supply chain dynamics. We will show an analogy between the IOBPCS family and
the Intel supply chain model to explore the underlying mechanisms of the dynamic properties in the
hybrid MTS-MTO system and propose corresponding mitigation strategies.
3. The Intel supply chain model
3.1 Model description

Gross assembly completion

AG pull AG Governs AG in pull mode
rate
A*G Desired AG push AG Governs AG in push mode
AN Net assembly completion rate S Actual shipments
A*N Desired AN S* Desired shipments
AWIP Assembly work in process SMAX Feasible shipments
*
AWIP* Desired AWIP WOI Desired weeks of Inventory =TOP + TSS
AWIPADJ AWIP adjustment WS Wafer starts =WS*
B Backlog WS* Desired wafer starts
B* Target backlog WSN* Desired net WS
BADJ B adjustment YD Die yield
D Actual order YL Line yield
Di* Desired die inflow YU Unit yield
DD* Desired delivery delay TDAdj Forecasting smoothing constant
Di Die completion rate TA Assembly time
DPW Die per wafer TB Time to adjust backlog
ED Long-term demand forecast TAWIP Time to correct AWIP discrepancy
ES Expected shipments TF Fabrication time
FG Gross fabrication rate TFGI Time to adjust FGI discrepancy
Finished goods inventory
FGI TFWIP Time to adjust FWIP discrepancy
stock
Information process time (Delay before
FGI* Target FGI TOP
shipments)
FGIADJ FGI adjustment TSAdj Shipping smoothing constant
FWIP Fabrication work in process TSS Safety stock coverage
FWIP* Desired FWIP FWIPADJ FWIP adjustment
IOBPCS Inventory and order-based production control system
VIOBPCS Variable inventory and order-based production control system
APIOBPCS Automatic pipeline and inventory and order-based production control system
APVIOBPCS Automatic pipeline and various inventory and order-based production control system
Table 2. Variables, constants and model descriptors used in the semiconductor supply chain model.
 B8 B5 B2
Available
capacity FGI availability
Wafer availability AWIP availability
+ Fabrication WIP + Assembly WIP Push AG + Finished Goods +
Actual Shipment
(FWIP) (AWIP) Net Inventory (FGI)
Wafer Net Min
Starts (WS) assembly Min
Production rate (Di)
Pull AG +Completion (AN) _ B1 Customer
+ FGI adjustment (FGIADJ) Desired + order
Shipment (S*) FGI pull
B7 B4 R1
_ + Actual Shipment (S)
_ Replenishment +
Adjust FWIP Adjust FGI +
Fab WIP AWIP Expected _ Actual
_ adjustment (TFWIP ) Adjustment (TAWIP ) Shipment Backlog level
B6 Order (D)
+ + (ES) (B)
Capacity Desired net B3
utilization Adjust AWIP assembly +
+ + + completion Assembly pull
Desired net rate (A*N )
Desired wafer production rate (WSN*) + Backlog
R2 +
Start (WS*) + Adjustment
+
(BADJ)
Production push
Forecast +
customer
demand (ED)

The result is give by the

Materials flow Min B Balancing loop R Reinforcing loop
minimum of two signals

Cumulative level
Information flow Smoothing representation
representation

Figure 2. Basic structure of the production-inventory based semiconductor supply chain system. Based on Gonçalves,
Hines, and Sterman (2005)

Figure 2 presents the basic material and information flows for the Intel supply chain including
fabrication, assembly and distribution (Gonçalves, Hines, and Sterman 2005). All the nomenclature
and model descriptors utilized in this paper are presented in Table 2. It should be noted that the
dynamic influence of customers’ response, i.e. the customers’ response to supply availability measured
by the fraction of order fulfilment, is not considered as we only focus on the effect of dynamic
production and inventory control in the Intel supply chain model. Here we give a brief introduction to
the supply chain operational design, while full details can be found in Gonçalves, Hines, and Sterman
(2005).
Specifically, there are two main manufacturing stages for microprocessor chip production from
a material flow perspective: fabrication and assembly. The polished disk-shaped silicon substrates
(wafers) as inputs are taken into a wafer fabrication facility, and through several complicated
sequences to produce fabricated wafers (composed of integrated circuits, i.e. ICs or dies). A vertical
cross-section of an integrated circuit reveals several layers formed during the fabrication process.
Lower layers include the critical electrical components (e.g. transistors, capacitors), which are
produced at the ’front-end’ of the fabrication process. Upper layers, produced at the ‘back-end’ of the
fabrication process, connect the electrical components to form circuits. In the second assembly phase,
the fabricated wafers are cut into dies and stored in the ADI warehouse to wait for the assembly process.
After passing assembly and test plants to ensure operability, the finished microprocessors are stored in
the FGI for customer orders. A three-stage supply chain, including fabrication, assembly and
distribution, is thereby created to represent the manufacturing process.
Regarding the information flow, the hybrid MTS-MTO information control strategy is
implemented. The downstream assembly and distribution systems are essentially the MTO mode in
which end customers’ orders pull the available microprocessors from FGI if there are sufficient FGI
and AWIP. The upstream wafer fabrication, however, is characterized by the MTS production style:
long-term demand forecasting and the adjustment from downstream AWIP and FWIP to determine the
desired wafer production rate.
The exogenous demand into the supply chain system begins when end customers’ demand
information is transmitted into the information system and tracked until it is shipped or cancelled. The
actual shipment, S, is determined by the minimum value between S* and SMAX. By design, the
distribution system operates as the MTO mode in which the S* is given by the ratio of B and DD*.
However, if insufficient FGI constrains S*, the distribution system will automatically switch to the
MTS mode to push all feasible FGI, which is estimated by FGI stock and TOP. As a result, those
backlogged orders directly pull product components from AWIP to increase the assembly order rate,
under the condition that the assembly system still performs the MTO-based production with enough
AWIP. The delivery delay experienced by external customers is increased in such scenarios, since
required orders cannot be fulfilled directly by FGI and it takes longer to assemble and distribute those
backlogged orders.
While shipments deplete FGI, the AN, defined by the AG and YU, replenishes it. AG is
determined by the minimum of pull AG and push AG signal. By design, pull AG is given by the desired
pull signal under the MTO operation in the assembly system, i.e. A*N adjusted by YU. AN* is determined
by the summation of the recent shipment, FGI adjustment and B adjustment. If all available AWIP still
constrains the assembly activities, the assembly system can only complete what are feasible and
thereby switch to the MTS production model, i.e. push AG, which is estimated by the ratio between
current AWIP and TA.
The upstream fabrication plant follows the MTS production strategy in which the produced
wafers are pushed into the ADI, the place where AWIP are stored until orders for specific product from
downstream assembly and distribution pull/push them depending on its availability. While AG depletes
AWIP, DI replenishes it. DI, measured in die per month, depends on FG (wafers per month), adjusted
by DPW and YD, i.e. the fraction of good die per wafer and YL to indicate the fraction of the good
fabricated wafers. For simplicity, a first order delay is utilized for modelling process. While FG depletes
FWIP, WS* replenishes it. The fab managers determine WS* based on gross WS and FWIPADJ. The
former is determined by D* required by assembly/test plants, which is based on a long-term demand
forecast (ED) and an adjustment from AWIP, while FWIPADJ depends on discrepancies between FWIP*
and FWIP adjusted by TFWIP. The capacity utilization (CU) is set based on ratio between WS* and
available capacity (K) operating at the normal capacity utilization level (CUN= 90%). The remaining
10% spare capacity is utilized for engineering purpose and to deal with manufacturing instability. For
a given D, K is determined by: , where MS (market share) is not considered in this
study. When WS* is larger than normal capacity utilization, Fab managers try to increase CUN and
thus the spare capacity for engineering is reduced. On the other hand, When WS* falls below the
normal CUN, capacity utilization will vary enough to exactly match WS*. However, field study
(Gonçalves, Hines, and Sterman 2005) showed that the managers prefer to build inventory by keeping
Fab running even when WS* falls below the normal capacity utilization. As the result, WS* can be
fully met by adjusting capacity utilization level and Fab managers prefer the ‘Lean-based’ production
to avoid machine shut down for the low capacity utilization scenario.
Based on this empirical information, we can separate the Intel supply chains into three
distinctive scenarios as follows:

 Fabrication MTS + Assembly MTO + Distribution MTO mode. Such a system is highly
desired since the customers’ orders can be fulfilled immediately by FGI (sufficient on-
hand FGI and AWIP inventory). The only waiting time for customers is the delivery delay,
which is assumed to be a first-order delay.
 Fabrication MTS + Assembly MTO + Distribution MTS mode. If FGI is insufficient for
customers’ orders, Intel can only ship what is feasible (SMAX) and transfer the
backlog/inventory signal into the assembly process to raise the assembly rate. In such a
condition, the assembly system still operates the MTO mode under the premise that there
are sufficient AWIP. The lead time for backlogged orders is increased to the summation of
the delivery delay and assembly delay.
 Fabrication MTS + Assembly MTS + Distribution MTS mode. If the assembly is also
constrained by the feasible AWIP level, the whole supply chain system will switch to a
 pure MTS mode. The customer orders cannot be fulfilled for a short time, due to the long
delay in fabrication production, and the lead time for backlogged orders is increased to the
summation of the delivery delay, assembly delay and fabrication delay.

It can be concluded that the Intel supply chain is a typical multi-product, multi-level production
environment where the final processors are produced through a series of sequences, from fabrication
to assembly and final shipments processing. Customer demand ultimately determines the
microprocessors’ production, and the whole supply chain system operates as a hybrid MTS-MTO in
which the actual shipment and assembly completions are driven by incoming demand orders, while
the upstream wafer production is influenced by long-term demand forecasting. However, the MTO
part will automatically switch to an MTS mode if there is no feasible AWIP/FGI in the assembly or
distribution system, and customers will experience the corresponding delay increase due to the switch
from MTO to MTS.

3.2 Extracting the hybrid MTS-MTO supply chain

Based on the causal loop diagram of Figure 2, and the detailed model description in Section
3.1, we developed the Intel supply chain model in a block diagram form, using continuous time domain,
Laplace s, as shown in Figure 3. In a recent publication, Naim et al. (2017) accomplished the same
resulting block diagram but in discrete time.

-+

B*
-+ B

S*
ED S
WOI* Min
SMAX
FGI*

BADJ Pull AG AG AN
+ A*N
+-
+ + FGI
- + + Min
FGIADj
0
Push AG

A*G

AWIP*

++ AWIPADJ + AWIP
-
0

DI
WS*N F*G FWIP* FWIP
D*I +-
+ -

FWIPADj WS
FG
+ +-
-
0
 Figure 3. Block diagram representation of the Intel supply chain

To analytically explore the underlying dynamic behaviour of the hybrid MTS-MTO

semiconductor supply chain systems, we simplify the block diagram through the following procedures,
following Wikner, Naim, and Towill (1992):

1. Transfer non-negative components into linear approximations.

Eliminating three non-negative nonlinear constraints by assuming the relevant variables
are never negative. Thus, non-negative constraints that restrict D*I, A*N, and WS are eliminated.

2. Supply chain echelon elimination

We assume that there is no distribution delay and that what is assembled into the FGI
can be directly fulfilled by external customer demand, that is, the distribution echelon is
eliminated. Thus, the backlog orders can be represented by negative FGI under the linear
assumption of Step 1, and the switch between S* and SMAX is eliminated. The whole model
now becomes a two-stage supply chain system.

3. Redundancies elimination
a. Given the assumption that the shipment made is equal to the demand, that is, S=D,
then B = DD D and B* = DD* D so that BADJ = 0
b. ED=ES
c. SMAX is redundant, given Step 2.

4. Collecting terms
Gross WS* is determined by the desired net wafer start rate adjusted by YL and in turn,
the desired wafer production rate is determined by D* in assembly, adjusted by the DPW and
the die yield YD, so we have following relationship:

Gross

To simplify the block diagram, we collect terms as follow:

a.
b.
c.

Since the linear model shown in Figure 4 is now considerably simpler than the original complex
supply chain, it can no longer be referred as the Intel supply chain, instead, the model is, from now on,
termed as a semiconductor hybrid MTS-MTO supply chain. One benefit of investigating the linear
system is that it enables the analytical tracing of supply chain dynamics. Given that, in reality,
semiconductor manufacturing suffers high capacity unevenness (Karabuk and Wu 2003) due to
reactive capacity adjustment driven by the dynamic behaviour, there is a need for managers to
proactively control the supply chain dynamics, and, especially, the bullwhip effect, via understanding
the root causes of such dynamic capacity requirements responses. This can be attained by assuming
linearity and using well-established linear control techniques to explore the impact of major control
policies on the dynamic behaviour. However, given that the simplification process and the linear
assumptions necessary for the analytical investigation may impact on the accuracy of responses and
on certain variable interactions, we will cross-check the analytical results (to be presented in Section
4) with numerical simulations of the nonlinear model (to be presented in Section 5) in order to enhance
the dynamic insights into the hybrid MTS-MTO supply chain model.

D
ED
WOI*

FGI*
FGIADj Pull AG AG AN FGI
+ A*N
+-
+
- + Min

Push AG

A*G

AWIP*
++ AWIPADJ + AWIP
-
D*I +-
+ -
DI
FWIP* FWIP

FWIPADj
FG
+ +-
- WS

Figure 4. Simplified block diagram for the hybrid MTS-MTO supply chain model.

As shown in Figure 4, the only nonlinearity left is the ‘Min’ function to govern the push/pull
downstream assembly activity, which we have deliberately maintained at this stage as it governs the
location of the decoupling point. If there is sufficient AWIP, i.e. push AG > pull AG, for customer orders
to pull chips from, then the whole system is fundamentally a hybrid MTS-MTO supply chain, i.e. the
MTS-based wafer fabrication and MTO-based assembly. Thus, AWIP is the CODP that separates the
upstream wafer production and downstream assembly activities. By contrast, if the AWIP is
insufficient to meet the pull signal, i.e. push AG < pull AG, all AWIP will be pushed into the assembly
plant to meet customer orders as soon as possible, and the whole system will automatically switch to
a pure MTS supply chain. As the main objective of this paper is to understand the underlying dynamic
properties of a hybrid MTS-MTO system, we focus exclusively on such a scenario.

4. Dynamic modelling and analysis of the semiconductor hybrid MTS-MTO supply chain
4.1 Modelling the hybrid MTS-MTO mode
As a result, the ‘Min’ function and push AG in Figure 4 are removed and the whole system now
is a typical hybrid MTS-MTO supply chain. We rearrange the structure to yield Figure 5 so as to draw
an analogy to the IOBPCS family. It can be seen that the MTS-MTO system consists of a VIOBPCS
(Edghill and Towill 1990) ordering rule in the downstream assembly stage and a structure similar to
the APVIOBPCS (Dejonckheere et al. 2003) ordering rule in the upstream fabrication.
The AWIP is the interface (CODP) connecting the fabrication and assembly production, i.e. the
AWIP is the finished stock point for the MTS fabrication, while it supplies raw materials for the MTO
assembly pulled by the customer ordering rate. For the downstream MTO system, represented by the
VIOBPCS, the only input is the customer demand signal. The block diagram also indicates that there
is an instantaneous assembly process that has a zero-yield loss for what is required for assembly, due
to the MTS-MTO condition that pull AG is always larger than push AG.
D Downstream MTO part
(Similar to VIOBPCS
archetype)

ED

FGI*
FGIADj + A*N AG AN
+
+
-
- +
FGI

Upstream MTS part

(Similar to the APVIOBPCS
A*G archetype)

AWIP*
+ + D*I WS FG -
+ + +
- +
Assembly
AWIP FWIPADj
-
+

FWIP* FWIP
+ -

Figure 5. Simplified block diagram for a semiconductor hybrid MTS-MTO mode.

As such, the desired rate (ordering rate), AN*, equals the net assembly complete rate (AN) as
follows:

Where

and
Towill

The upstream MTS fabrication system is similar to the APVIOBPCS replenishment rule that
includes inventory feedback correction (AWIP), work-in-process feedback correction (FWIP) and
feedforward forecasting compensation (ED). There are two inputs in such a system including demand
from the MTO system and the end customer, and feedforward forecasting, i.e. ED(t), is based on the
end customer demand (D). Therefore, the ordering rate for each replenishment cycle is given by:

and AWIPADJ(t) is determined by the fraction of difference between the desired assembly pull
level and actual AWIP level, which equals:
 FWIPADJ(t) is determined by a fraction of the difference between the desired inflow FWIP and
the actual FWIP as follows:

It should be noted that the safety stock levels, AWIP and FGI*, and desired FWIP* are based
on constant gains, WOI*, TA and K2/YU, that need to be set. Hence, there is an opportunity to further
explore the impact of setting such levels, which give more insight on the overall dynamic behaviour
of the hybrid MTS-MTO system. e.g. see Manary and Willems (2008) method to address the issue of
systematically biased forecast experienced by the Intel supply chain. However, it is beyond the scope
of this paper to investigate the impact of parameter variation on the dynamics of the hybrid MTS-MTO
semiconductor supply chain.
In order to allow us to benchmark the dynamic behaviour of the upstream MTS representation
with an exact APVIOBPCS, we redraw the block diagram in Figure 5 to represent the exact
APVIOBPCS system as shown in Figure 6.

D Downstream MTO part

(VIOBPCS archetype)

ED

FGI*
FGIADj + A*N AG AN
+
+
-
- +
FGI

Upstream MTS part

( the APVIOBPCS archetype)
A*G

D*I
AWIP* AWIPADJ
+ + WS FG -
+ +
- +
Assembly
AWIP FWIPADj
-
+

FWIP* FWIP
+ -

Figure 6. The APVIOBPCS-based hybrid MTO-MTS in block diagram form

From Figures 5 and 6, we may observe that, unlike the traditional APVIOBPCS ordering rule
that has only one input, i.e. demand from the MTO system, there are two inputs utilized for the wafer
production rate in the semiconductor MTS system: 1) demand from the next-level supply chain echelon;
and 2) demand from the end customer order. Such a structure is, fundamentally, a material requirement
planning (MRP) system, while the APVIOBPCS has been defined as a ‘Reorder system’ (Popplewell
and Bonney 1987). Table 3 summarizes the difference between the two ordering rules for the MTS
system.

Type of system Targeted Inventory Targeted WIP Feedforward

(the MTS) (feedback loop) (feedback loop) forecasting loop
Semiconductor As a function of demand As a function of the summation Based on final
MRP system from the ordering rate at of inventory correction (AWIP) customer demand (D)
assembly production (AN*) and demand from end customer
(D)
Reorder system As a function of demand As a function of demand from Based on demand from
(APVIOBPCS) from the ordering rate at the ordering rate at assembly the ordering rate at
assembly production (AN*) production (AN*) assembly production
(AN*)

Table 3. The comparison of system structure between the semiconductor MRP and APVIOBPCS systems.

In summary, the final stylised hybrid MTS-MTO structure consists of two major ordering rules
under the assumption that the CODP inventory are always available for end customers’ orders.
Downstream of the CODP is the VIOBPCS ordering rule with negligible lead time; while the upstream
MTS structure is similar to the APVIOBPCS, but there are some differences regarding the settings of
the targeted WIP feedback loop as well as the feedforward forecasting loop. Using a control
engineering approach, we now investigate the underlying dynamic behaviour of the semiconductor
hybrid supply chain system.

4.2 Transfer function analysis

As we focus on the dynamic behaviour of the inventory and order rate in responding to the
external demand signal under the hybrid MTS-MTO supply chains (Figure 5), the corresponding
transfer functions, downstream FGI, A*N in relation to the demand (D), can be derived based on the
following procedures:

 Substitute Equation (2) into (1);

 Substitute Laplace domain of ED in relation to D, i.e. , into Equation (1);
 Substitute Laplace domain of FGI in relation to A*N, i.e. , into Equation (2) and
then Substitute Equation (2) into (1).

We now have the transfer function of A*N in relation to D:

Substitute Equation (7) into , the transfer function of FGI can be derived thus:

Similarly, the upstream WS and AWIP in relation to D can be derived by the following steps:

 Substitute Equation (5) and (6) into (4) in Laplace form to obtain:
 Where

 Substitute Equation (7), (10), (11) and (12) into (9)

Now we can obtain the transfer function of WS in relation to D as follows:

Substitute Equation (7) and (13) into (11), we can obtain the transfer function of AWIP in
relation to D:

The transfer function represents the dynamic properties of the system. In particular, the
characteristic equation, defined by equating the denominator of overall transfer function to zero, can
be used to find poles (roots), which give an initial understanding of the underlying dynamic mechanism
of the semiconductor hybrid MTS-MTO system including system stability and unforced system
dynamic property (i.e. natural frequency and damping ratio). Stability is a fundamental property of a
supply chain system. From the linear system perspective, the system is stable if the trajectory will
eventually return to an equilibrium point irrelevant to the initial condition, while an infinity trajectory
is presented if the system is unstable (Wang, Disney, and Wang 2012).Thus, the system response to
any change in an input (demand) will result in uncontrollably increasing oscillations in the supply
chain (Disney and Towill 2002). A system also has critical stability when it is located at the edge of
the stability boundary, and system oscillations are regular and infinite for such situation. More details
of supply chain stability can be found in Riddalls and Bennett (2002), Warburton et al. (2004), Sipahi
and Delice (2010). Regarding the unforced system dynamic property, natural frequency ( )
determines how fast the system oscillates during the transient response and can be used to indicate the
system’s speed to reach the steady state condition for responding external demand signal, e.g. the
inventory recovery speed. Damping ratio ( , on the other hand, describes how the system’s oscillatory
behaviour (i.e. variability) decays with time, and can be perceived as initial insight of the system’s
unforced dynamic performance; for instance, the extent to which the order rate and inventory will
oscillate with time.
We now focus on the characteristics equations. By rewriting the denominator of Equation (7),
(8), (13) and (14) as Equation (15), it can be seen that the MTO is characterized by a second-order
system, while a fourth-order polynomial describes the MTS system:

Also, there is a second-order polynomial, , in the denominator of all

transfer functions, which confirms that the dynamic property of the MTO system is not influenced by
the MTS system, while the dynamic performance of the MTS system can be partially manipulated by
the MTO system under the hybrid MTS-MTO mode.
We now turn to the analysis of Initial Value Theorem (IVT) and Final Value Theorem (FVT).
The IVT is a useful tool to cross-check mathematically the correctness of a transfer function and guide
the appropriate initial condition required by a simulation. The FVT is useful to understand the steady
state value of the dynamic response of a transfer function and can help verify the simulation. Equation
16 presents the initial and final values of FGI, AN*, WS and AWIP in responding to a unit step input
for the semiconductor hybrid MTS-MTO system.

As expected, the initial values of FGI, AN*, AWIP, FWIP and WS are zero; similar to the results
obtained by John, Naim and Towill (1994). Regarding the final value, the ordering rate (A*N) of the
MTO system is unity and the steady state level of the FGI is WOI* as it is a function of the averaged
demand. The final value of ordering rate (WS) for the upstream MTS system is, as expected, a system
constant value , and the final value of AWIP is determined by the coefficient (the targeted
inventory level in the APVIOBPCS). Since the downstream MTO system is not influenced by the
upstream MTS system, due to the assumption of infinite AWIP availability to maintain the MTO
assembly while the dynamic behaviour of MTS is influenced by the upstream MTO, we analyse the
dynamic properties of the MTO and MTS systems separately.

4.3 Characteristic equation analysis of the MTO system

Since the transfer function of the MTO part is a second-order system, its associated dynamic
properties are defined by and determined by the characteristic equation. Hence, we obtain
and as follows:

Based on Equation 17, both and are determined by the control parameters TDAdj and TFGI.
The natural frequency decreases as the values of TDAdj and TFGI increase, leading to a slower dynamic
response and recovery to the steady state conditions for the MTO system. To visualize the relationship
between and TDAdj and TFGI, we rewrite Equation 17 as 18:

When TDAdj = TFGI, always assumes the same value hhen either TDAdj or TFGI
increases, increases further decreasing the number of oscillations in responding to external
demand but making the system slow The important message here is that for all positive
values of TDAdj and TFGI, which means that the system always produces over-damped behaviour and is
guaranteed to be stable. This is important because the system is permitted to be stable and robust for
any choice of positive decision-making parameters. Furthermore, we cannot achieve both objectives
of the rapid inventory recovery (natural frequency) and low level of bullwhip (i.e. maximum overshoot)
determined by the damping ratio. This trade-off has also been confirmed mathematically by Towill
(1982).

4.4. Unit step response of the MTO system

The unit step input is utilized to assess the dynamic behaviour of the semiconductor hybrid
MTS-MTO system. The step as an input source is well documented (Towill 1970) in general control
theory for exploring the system’s capacity to respond to sudden but sustained change. Moreover, step
change as the input is easily visualized and its response can be easily interpreted (John, Naim, and
Towill 1994). Furthermore, the step increases give rich information for the dynamic behaviour of the
system (Coyle 1977). From the supply chain point of view, the step demand can be regarded as the
early stage of a new product or the opening of a new sales outlet (Zhou and Disney 2006), which fits
the customer demand condition in the semiconductor industry characterized by a short life cycle with
a corresponding sudden change in demand during the release of new products.
Due to the analogy between the VIOBPCS and the MTS mode of the semiconductor hybrid
supply chain system, the set of parameters utilized is as suggested by Edghill and Towill (1990) with
4 units of assembly lead time (TA=4). Based on the transfer functions of the MTO system, i.e.
Equations (11) and (12), the value of required system parameters for simulation are thereby (weeks):
TDAdj =8, TFGI=4 and WOI*=5
 Figure 7. The impact of T and T for FGI and A unit response.
Figure 7 demonstrates the impact of TDAdj and TFGI for the unit step response of the FGI and
A The solid line represents the recommended settings utilized in the VIOBPCS archetype. There is
always an initial drop for the FGI response due to the transient response of a unit step increase in
demand, and the absolute FGI drop value can thereby be utilized for setting initial stock levels to
maintain supply to the MTO system. When TFGI increases, the FGI response experiences a larger initial
drop with a longer setting time, while the A*N has a shorter setting time at the expense of higher peak
level. Similarly, a larger undershot and longer recovery time of the FGI response are observed when
the value of TDAdj increases, while the A*N experiences less bullwhip at the expense of a longer settling
time.
To summarise, the downstream MTO assembly system always produces over-damped dynamic
behaviour and such a system is guaranteed to be stable and robust, although there is an overshoot for
AN transient response due to the effect of the numerator of transfer functions. Bullwhip results from
TDAdj and TFGI, which confirms the fact that forecasting (Dejonckheere et al. 2002) and feedback loops
(Lee, Padmanabhan, and Whang 1997) are the major sources of bullwhip generation, even when the
lead time is negligible. In particular, TDAdj places a major emphasis on the bullwhip level, while TFGI
has a major impact on the FGI variance. This result also provides evidence that bullwhip is mainly
caused by the feedforward compensation, instead of the feedback loop/production delay usually
suggested. Although this phase advance/predictive component (Truxal and Weinberg 1955) in the
hardware control engineering field has the advantage of ordering in advance to ensure stock availability,
some solutions such as more sophisticated forecasting algorithms (Dejonckheere et al. 2002) must be
implemented to reduce the bullwhip level.

4.5 Characteristic equations analysis of the MTS system

Based on Equations (15), the MTS system is characterized as a fourth-order polynomial that
can be rewritten as the product of two second-order polynomials. As the second-order polynomial, i.e.
was already analysed in the MTO system, we derive the natural frequency
and damping ratio for the other second-order polynomial as follows:

For a fixed TF (physical fabrication lead time), and are determined by TAWIP and TFWIP.
The system response will become slower (smaller value of ) as TAWIP and TFWIP increase. However,
TAWIP and TFWIP have a reverse impact on . The system will be more oscillatory as TFWIP increases or
TAWIP decreases. It should be noted that TAWIP has a major influence on the damping ratio compared to
TFWIP, which means the CODP inventory policy plays a major role in the system’s dynamic behaviour.
To further understand the dynamic properties of the MTS system, including transient response
and stability, we derive the four poles based on Equation (15) as follows:

There is no imaginary part for the roots of the first and second polynomials ( and ), and
thereby oscillatory behaviour cannot be generated. For and , the roots can be real, complex or
purely imaginary and the real poles can also be positive, negative or repeated, influencing the transient
response as well as the stability condition. We plot the different results of roots based on different fixed
values ranging from 1 unit to 4 units as shown in Figure 8.
 Figure 8. Real, complex and imaginary region of R3 and R4 based on different TF.

Figure 8 illustrates that the roots are positive for the region between the line of purely
imaginary roots and TFWIP = 0, thus, the pair choice of TAWIP and TFWIP in this area will lead to an
unstable system. Also, we consider the impact of negative FWIP feedback controller (T FWIP) on the
results of the roots, although, conventionally, it is assumed to be a positive value range. The negative
TFWIP has been investigated in the case of a uniformed and irrational replenishment rule design (Wang,
Disney, and Wang 2012, 2014). Based on Figure 9, the roots will become purely imaginary if the real
part of the roots is zero (i.e. TFWIP= -TF). Furthermore, the purely imaginary roots are the critically
stable point; as such, the system response will be sustainably oscillatory.
Although the transient response of the fourth-order system is multifaceted, determined by the
dominant pole(s) that is/are closest to the origin of the s plane, i.e. the combination of different control
policies, the result in Figure 8 gives a qualitative understanding of the system’s dynamic properties,
i.e. whether stable or unstable, for different parameter choices. For instance, we can specify the range
of TFWIP and TAWIP to generate real poles, i.e. a ‘good’ system dynamic design without generating
oscillations. As a result, semiconductor companies may benefit from associated cost reduction by
improved supply chain dynamics performance. In addition, the real poles region becomes smaller as
fabrication lead time, TF, increases, which means that the system is more likely to generate oscillatory
behaviour based on different choices of decision parameter settings. Managers thus need to be aware
that their upstream MTS systems are more likely to be oscillatory under their control policies if
fabrication lead times become longer. Finally, we can conclude that such a semiconductor hybrid MTS-
MTO system is stable for all positive decision parameter choices (TFWIP, TAWIP, TFGI, TDAdj).

4.6 Unit step response of the MTS system

To understand the impact of four system policies (TFWIP, TAWIP, TFGI, TDAdj) in influencing the
transient response of the hybrid MTS-MTO system, we conduct a step response analysis through the
initial settings suggested by John, Naim, and Towill (1994), i.e. for
the dynamic performance of the MTS system. The recommended settings of both VIOBPCS and
APVIOBPCS will be utilized as the initial design to determine whether such parameter settings can
still produce ‘good’ dynamic performance in the hybrid environment. The system’s constant
parameters including K1, K2, K3 and Yu will be discarded, as they do not influence the system’s
dynamic behaviour.
We assume that the lead time ratio between assembly and fabrication is 1:2 (i.e. 4 and 8 for
assembly and fabrication) to represent the long-term upstream fabrication and relatively short time for
the customized assembly. Thus, the initial setting is as follows (weeks):
Figure 9. The effect of decision policies for the AWIP and WS in responding to unit step increases.
 Figure 9 shows the impact of TFWIP, TAWIP, TFGI and TDAdj on the dynamic behaviour of the WS
and AWIP in the MTS mode. The solid line represents the recommended settings used in the VIOBPCS
and APIOBPCS archetypes. Compared to the downstream MTO mode, the bullwhip and inventory
variance are more significant in the upstream MTS mode, due to the fact that the dynamic behaviour
is amplified from the end customer to the far position of the entire supply chain (e.g. manufacturer).
The AWIP always experiences an initial drop in responding to unit step input, as the AWIP must meet
the downstream customer MTO signal during the transient period to maintain the hybrid MTS-MTO
mode. The AWIP recovers to the desired level with a gradual increase in the fabrication production
complete rate to match the unit step demand increase. The absolute decline level is helpful to indicate
the safety inventory required to maintain the hybrid mode during the transient period.
Based on Figure 9 and Table 5, an increase in TDAdj leads to a longer peak time and setting time,
but less oscillation of the AWIP. Moreover, an increase in TDAdj slightly reduces the peak level of the
AWIP. It should be noted that the AWIP exhibits oscillatory behaviour for small values of TDAdj in
responding unit step increase, due to the long-term fabrication delay (TF) and the amplified pull signal
downstream (AN) as the input of the MTS part. Similarly, the WS also experiences less bullwhip and
fewer oscillations as TDAdj increases at the expense of a longer setting time. Similarly, an increasing
TFGI reduces the overshot and undershot of the AWIP compromised by a slightly longer setting time.
However, for a sufficiently long FGI correction time (large TFGI), there is no system overshot for the
AWIP with a much shorter setting time. The WS experienced a high bullwhip level and more oscillation
under small values of TFGI.
Regarding the decision parameters in the upstream system, TAWIP significantly influences the
dynamic response of the AWIP and WS. An increase in TAWIP dramatically increases the undershot
(also peak time) and setting time of the AWIP, while the WS has less bullwhip, oscillations and a
shorter setting time. In particular, a small TAWIP introduces extra oscillatory behaviour in response to
the AWIP and WS, due to the feedback loop control and long production delay. An increase in TFWIP
damages the dynamic performance of the FWIP by producing more undershot and oscillations with a
longer setting time. Similarly, the WS response has more oscillations and a longer setting time at the
expense of less bullwhip as TFWIP increases. Since the target FWIP is the summation of ED and
AWIPADJ (AWIP feedback loop has been included for AWIPADJ), the long correction time for the
feedback FWIP loop will further amplify the effect of the AWIP feedback loop by introducing extra
dynamic behaviour for the AWIP and WS, which damaging their dynamic performance via introducing
more oscillations. Furthermore, based on Figure 9 and Table 3, the recommended settings in the
APIOBPCS and VIOBPCS can still be utilized in the hybrid MTS-MTO supply chain to yield a ‘good’
dynamic response when considering the trade-off between bullwhip and inventory recovery. We
summarize four decision parameters’ impact on the hybrid MTS-MTO step response by increasing
their value in Table 4:
Decision AWIP WS FGI AN
parameters
p tp ts p tp ts p tp ts p tp ts
TsAdj 0

TFGI

TAWIP 0 0 0 0 0 0

TFWIP 0 0 0 0 0 0

Table 4. Summary of the system response by increasing the value of decision parameters (p: peak level, t p: time for peak
level, ts: setting time, : better performance. : worse performance, 0: no influence, : from worse to better performance
due to the extra oscillations).
It can be concluded that maintaining the hybrid MTO-MTS system in the semiconductor
industry is highly desirable since customer orders can be fulfilled immediately. Feedforward
forecasting compensation and three feedback correction loops (FGI, AWIP, FWIP) have an impact on
the bullwhip level. In particular, the CODP inventory policy (TAWIP) and the forecasting policy (TDAdj)
significantly influence the bullwhip level; TAWIP also plays a major rule in the system’s oscillatory
behaviour. Thus, managers should carefully tune TAWIP to balance the benefit between the cost of
holding CODP inventory and the cost of supply chain dynamics. Moreover, practitioners should
consider the choice of TFGI to balance the levels of two safety stock points (AWIP and FGI), as such a
policy has a reverse influence on the AWIP and FGI. Finally, the recommended settings in the
APVIOBPCS and VIOBPCS are still ‘good’ in the semiconductor hybrid system, although there are
some differences between the APVIOBPCS-based reorder system and the MRP-based replenishment
rule. Furthermore, the dynamic response of AN and WS, e.g. rising time, peak level and setting time,
give useful guidance for benchmarking the results derived from the nonlinear dynamic system to set
an optimal capacity in the nonlinear system, which may balance the cost of bullwhip and inventory
variance in responding to a sudden but sustained change in demand.

5. Simulation enhancement
Although the analytical results derived from the linear system above offer deep insights into
the system dynamic behaviour of a semiconductor MTS-MTO supply chain, linear assumptions are
often criticized for being incapable of capturing nonlinear characteristics of the real supply chain
system with resources constraints (e.g. capacity, non-negative order constrains) (Lin et al. 2016). To
enhance the qualitative insights obtained from the linear analysis, we incorporate the nonlinearities to
[0, Capacity limit]

represent the capacity, as a CLIP function ( ) and non-negativities in the hybrid MTS-MTO
model of Figure 5. It should be noted that there are a number of other capacity forms that can be used
to represent the capacitated semiconductor fabrication environment. e.g. see Orcun, Uzsoy and Kempf
(2006)’s exploration of the dynamic behaviour of Clearing Function (CF) based capacity models in a
simple capacitated production system.
The hybrid mode is still assumed to be in operation but in a resource constrained environment,
reflected in the block diagram representation shown in Figure 10. We should note that the CLIP
function is an addition that is not in the Gonçalves, Hines, and Sterman (2005)’s representation.

ED D Downstream MTO part

(Similar to VIOBPCS
archetype)

FGI* A*N AG AN
+ FGIADj + -
- + +
FGI
0

Upstream MTS part

(Similar to the APVIOBPCS
A*G archetype)
D*I [0, Capacity limit]

AWIP* AWIPADJ WS
+ FG
+ + -
++ +
-
0 FWIPADJ Assembly
AWIP
-
+

FWIP* FWIP
+ -

Figure 10. The semiconductor hybrid MTS-MTO supply chain in the nonlinear block diagram form.
Figure 11. WS and AWIP response for a step demand increase in the nonlinear hybrid MTS-MTO settings
 Similar to the linear system analysis, a step input is utilized and all system and control policies
settings remain the same. Capacity limit in the MTS part is set as 50% larger than the step demand (i.e.
1.5), since on average manufacturing capacity has to be larger than required demand to keep the system
stable. Figure 11 presents the impact of four system control policies (TDAdj, TFGI, TAWIP and TFWIP) on
the dynamic behaviour of the MTS part in the hybrid mode. The solid line represents the recommended
settings in the original APVIOBPCS and VIOBPCS archetypes, although it is not necessary to be
‘optimal’ in the nonlinear environment depending on the specific trade-offs design between inventory
and capacity. It should be noted that the corresponding control policies assessment for the MTO part
independent of the MTS part is not reported here, due to the little dynamic impact of non-negative
nonlinearity in responding to a step increase in demand, i.e. the same dynamic behaviour is observed
in the nonlinear MTO system
In general, the simulation of a nonlinear hybrid MTS-MTO system shows that the insights
obtained from the linearized analytical results are correct. The increase of policies in MTO part, i.e.
TFGI and TDAdj, negatively influences the dynamic performance of the CODP inventory (AWIP) by
introducing more undershoot and longer setting time, while the better dynamic responses of WS are
found with fewer oscillations and fast recovery speed. However, as expected, comparing the linear
results (Figure 9) under the same control policy settings, the step increase in demand gives higher
initial drop of AWIP and slower recovery speed of AWIP and WS in the nonlinear environment. This
is because more CODP inventory (AWIP) is needed and longer recovery time is influenced by the
period when the manufacturing rate hits the capacity limit. Furthermore, TAWIP significantly influences
the dynamic performance of the MTS part in the nonlinear hybrid system in terms of oscillations and
recovery speed. The WIP correction policy in the MTS part, that is TFWIP, as expected, reported the
same qualitative insights obtained from the linear system in which an increase in T FWIP lead to the
worse dynamic behaviour of AWIP and WS by introducing more undershoot and oscillations. The
whole hybrid MTS-MTO system experiences a significant reduction of bullwhip level (WS) at the
expense of more AWIP variability in a capacitated based nonlinear system, in comparison with results
obtained from the linear system.

6. Discussion and Conclusion

In this paper, we explored analytically the dynamic properties of a hybrid MTS-MTO supply chain
system within the context of semiconductor industry. We used the supply chain model, empirically
reported by Gonçalves, Hines, and Sterman (2005), as a benchmark model, to extract the MTS-MTO
model and explore the underlying properties of such hybrid system in the semiconductor production
environment. By utilizing control engineering techniques and the well-known IOBPCS family of
archetypes, we addressed the limitations of Gonçalves, Hines, and Sterman's (2005) simulation work
that lacks the analytical results and guidance for practitioners about the underlying root causes of
supply chain dynamics in a hybrid MTS-MTO supply chain environment. The summarized findings
and corresponding managerial implications are presented in Table 5:

System Findings Corresponding managerial implications

1. Quick forecasting smoothing and
1. TsAdj and TFGI have a reverse impact inventory error correction lead to rapid
on . system’s recovery to the steady state
2. 1 for all positive value of TsAdj and condition.
Linear MTO TFGI. 2. The MTO system always produces over-
damped behaviour without oscillations.
The MTO system is permitted to be stable
The real roots are always negative for
Stability and robust for any forecasting methods and a
positive value of TsAdj and TFGI.
positive value of inventory correction policy.

TsAdj plays a major role in the AN
Dynamic
response, while TFGI has a major impact
response
on the FGI dynamical behaviour.
Nonlinear Dynamic The same insights from the linear MTO
MTO response system are confirmed.
1. TAWIP and TFWIP have a reverse impact
on . conditions.
2.TAWIP has a major influence on .
The real roots are always negative for
Stability
positive value of TFWIP and TAWIP.
Linear MTS
1. TsAdj, TFGI, TAWIP and TFWIP influence
the bullwhip effect. However, T sAdj and
Dynamic TAWIP play major role for bullwhip
response level.
2. TAWIP is the key factor for system
oscillations.
The same results regarding the impact
of control policies on the dynamic
behaviour can be confirmed in the
Dynamic nonlinear MTS system. However, the
Nonlinear MTS
response introduction of nonlinearities can
reduce the bullwhip effect at the
expense of increasing CODP inventory
variability.
Bullwhip is mainly caused by the feedforward
compensation in the semiconductor MTO
supply chains: A careful compromise
between advance stock availability and
bullwhip effect should be considered.
The same managerial implications are
obtained.
1. Quick CODP (AWIP) and FWIP
inventory error correction leads to the
system’s rapid recovery to steady state
2. The CODP (AWIP) inventory policy plays
a major rule in the system’s dynamic
behaviour.
The MTS system is guaranteed to be stable
for any positive choice of the AWIP and
FWIP inventory correction policies.
1. The CODP inventory error correction and
forecasting smoothing policy should be
carefully tuned, due to their major influence
on bullwhip level in the MTS system.
2. The trade-off between the cost of bullwhip
(e.g. capacity ramp up/down, labour hiring
and firing) and the benefit of implementing
CODP strategy should be considered, due to
the system’s oscillations are sensitive to the
CODP policy settings.
Beside the managerial insights gained from
the linear analysis, the impact of capacity
constraint should be considered for trade-
offs design between the CODP inventory
and capacity utilization when the hybrid
MTS-MTO production strategy is adopted.

Table 5. Summary of the findings and managerial implications for the semiconductor hybrid MTS-MTO supply chains.
Source: the authors.

To answer research question 1, we gained deeper insights into the dynamic properties of the
hybrid MTS-MTO supply chain system by simplifying and linearizing the original complex dynamic
model, including developing the block diagram form, removing nonlinearities and redundancies and
eliminating one echelon of the supply chain system. Thus, it is possible to extract the scenario of the
linear hybrid MTS-MTO and implement a linear control engineering approach to analyse its
fundamental dynamic properties. Although the simplification method is based on the semiconductor
supply chain system, a similar approach can be applied to a board production-inventory based
manufacturing system.
Regarding research question 2, through control engineering approaches, including Laplace
transform, characteristic equations and the unit step response analysis, we reveal the fact that
feedforward forecasting compensation and the CODP inventory correction policy play a major rule in
the bullwhip effect in the semiconductor hybrid MTS-MTO system, instead of the production
delay/feedback loop usually claimed in practice. Also, semiconductor managers may need to
cautiously consider the balance between the cost of keeping an adequate CODP inventory to maintain
the mode of MTS-MTO and the cost of supply chain dynamics, due to the fact that the policies’ settings
in the CODP point are significantly sensitive to the inventory variance and bullwhip level. This finding
is helpful for practitioners to carefully consider relevant trade-offs when designing their hybrid MTS-
MTO system in the semiconductor industry.
Simulation analysis is also conducted based on a capacitated nonlinear hybrid semiconductor
system to enhance qualitative insights derived from the linear analysis. The results show that the
analytical results are robust in a nonlinear environment with capacity and non-negative order
constrains. However, in the capacitated setting the bullwhip level is reduced at the expense of
damaging CODP inventory performance (i.e. increased variability) when compared with the linear
system under same policy settings. This phenomenon is well-recognized in literature (Cannella,
Ciancimino, and Marquez 2008; Nepal, Murat, and Chinnam 2012; Lin, Jiang, and Wang 2014; Ponte
et al. 2017). Furthermore, this paper also contributes to the analysis of supply chain dynamics in the
semiconductor industry by uncovering the underlying mechanism of the bullwhip effect and
comparing the dynamic performance differences between two different ordering archetypes. While we
have undertaken some analysis of varying decision parameters with respect to proportional feedback
controllers, i.e. TFWIP, TAWIP, TFGI, TDAdj, further dynamic analysis needs to be undertaken such as
determining the impact of feedforward gains i.e. WOI*, TA and K2/YU.
It should be acknowledged that the IOBPCS-based production control frameworks (linear or
nonlinear representations) may not be capable to capture the nonlinear increase of cycle time in the
semiconductor fab production at high resources utilization condition (Orcun, Uzsoy and Kempf 2006),
due to the lead time modelling approaches, i.e. first order/third order delay under the fixed mean lead
time assumption. To be more specific, with the increase of WIP level, longer time is required for the
semiconductor fab system to transform releases into output, resulted by the nonlinear increase of cycle
time in both mean and variability, which may lead to the different dynamic behaviour under the high
capacity utilization level comparing the corresponding response of the IOBPCS family. Although a
similar behaviour can be obtained from IOBPCS-based and non IOBPCS-based systems (e.g. the
adoption of the clearing functions) at low utilization level, the exploration of alternative lead
time/capacity modelling approaches in the IOBPCS family, e.g. Clearing Function based capacity
models (Orcun, Uzsoy and Kempf 2006), in representing a more realistic semiconductor fab WIP
congestion condition is strongly recommended for the future study.
Nevertheless, the linear analysis in this paper still offers robust and traceable insights for the
effect of fundamental system structures (feedback, feedforward) and the corresponding control policies
on the dynamic behaviour of the hybrid MTS-MTO system, which contribute to the analytical guidance
of supply chain system design in the context of the semiconductor industry. The analytical stability
region map gives a basic framework for examining stability condition in both linear and nonlinear
environment. The well-established results derived from the linear system, bullwhip level, fill rate and
the corresponding economic implications, for example, can be also used as the indicators to compare
the nonlinear dynamic results. A good example is Ponte et al.'s (2017) method to set optimal capacity
based on the benchmark of the well-established linear analysis results in a capacitated production
environment. Although it is out of this paper’s scope, we suggest this as a future research direction
based on such a linearized model.
There are several other future research opportunities based on this study. First, the application
of nonlinear control engineering approaches should be considered by researchers, in particular for the
context of the semiconductor supply chains to guide practitioners design and improve their systems,
although common nonlinearities presented in the IOBPCS ordering family already be analytically
traced by utilizing such methods (Spiegler, Naim, et al. 2016; Spiegler, Potter, et al. 2016; Spiegler
and Naim 2017). In addition, this study is limited to the analysis of the hybrid MTS-MTO system only
and ignores cases in which the whole supply chain automatically switches to the pure MTS system if
there is no feasible FGI/AWIP in the semiconductor production context. Future researchers can
contribute to these areas by utilizing more advanced analytical methods and analysing the switch
between different modes within supply chain systems. It should be noted that the MTS part we derived
from the original supply chain model is a linear-based MRP system, thus it can easily be extended to
accommodate capacity constraints (Mönch, Fowler, and Mason 2013), to represent more realistic
semiconductor manufacturing scenarios. This also gives rise to a possible future research opportunity
to investigate the dynamic behaviour of capacitated semiconductor supply chains in responding to
various fluctuation demands.

Reference
B N ME K A CONWIP
I J P R
B M J N CS T
J O M
B A O H LL ‘ P X
I
C S E C A C M C A
I J S P M
C P S G M
I J P R
C J D M A ATO I
J P R
C Y F S MD T
I J P R
C C F Y J C J T P M
I J P E

C M L L P H E
C ‘G M C J W S A L
D J S MD M ‘L D ‘T T
I J P E
D J S MD M ‘L D ‘T M
A E J O R
D J S MD M ‘L D ‘T T
A E J
O R
D S M D ‘T ‘ DH W O
I J P E

D S M D ‘ T A
I J P R

D G H S N
E J O R
E J D‘ T A E
C P E
F L L M A
P W S C
G N J ) A
I J P R
G P J H J S T
S
G H J L F M E J O
R
G A C P ‘ EM G A
I J P E
H T P H LL J JN T
N Y S
H E B
I J P R
H Y B J F C ‘ H
 I J P R
J S M MN D ‘T D WIP
I J M S D
K S SD W C
O R
K JI SH K P
I J A L
K S H J WF D LS M EP I
I J P
R
L H L V P S W I T
M S
L Y H S J S
I
L Y H S C B L K H K S
P P C
L Q D S M ‘ M‘P I J
P ‘
L D C K M L A‘ ‘ M ‘ S P
I J P R
L J M M N L P J G T
I J
P E I
L W ) J L W M
I J P R
M M P S PW S I
I
M L J W F S J M P P C S W
F F B S V N Y
N MM S VL W J T D‘ I
E J O R

N J B
N B
N N
O J
M M N D B L
I J P E
A M ‘ B C T
I J P E
S CONTROL SYSTEMS ENGINEERING J W S
S I J P E

O S ‘ U T
I P P I E E S
O S ‘ U K K U
T W
P B X W D F S MD E
I J P R
P K M CB T
I J P
R
Q C A I S S BG I
IEEE T S M
‘ A PJ WF WM C O MA A M B B S

I J P E
‘ CE SB T I J P R

S S E SG J WF K GK A
I C I
S H P P C D T C T K D
A C O R
S J KF G S M C E A L E L
S H A O E
J E S
S ‘ I I D S
I J P E
S V LM M MN I
E J O R
S V LM M MN D ‘T J W A
E J O R

S V LM M MN J W A
I J P R
S V LM A T P MM N D ‘ T T
I J
P R
S J D M M
M S
S Y D L S J W F E S G S
C J O M
T C )J ‘ T HH M AC A
M S
T D ‘ D I
J P R
T D ‘ T L I
T D‘ E I J P R

T J G LW A P T
V J J ‘ S B P A P E
I J P R
W W D E ‘ M
IEEE T C S T
W W D E‘ K GK M
I J P E
W X S M D J W S
E J O R
W X S MD J W E
I J P E
W ‘ DH S MD D ‘T J PE H F
I J P R
W J M N D T T
J S E
) C B J F C ‘ C
I J P R
) L S MD B O S

) L S D D ‘ T A
I J P E
) L M M N S M D T
I J
P E
) L M MN O T D ‘T D
K O